Evaldas Vaiciukynas

An Integrated Approach to Analysis of Phytoplankton Images

The main objective of this paper is detection, recognition, and abundance estimation of objects representing the Prorocentrum minimum (Pavillard) Schiller (P. minimum) species in phytoplankton images. The species is known to cause harmful blooms in many estuarine and coastal environments. The proposed technique for solving the task exploits images of two types, namely, obtained using light and fluorescence microscopy. Various image preprocessing techniques are applied to extract a variety of features characterizing P. minimum cells and cell contours. Relevant feature subsets are then selected and used in support vector machine (SVM) as well as random forest (RF) classifiers to distinguish between P. minimum cells and other objects. To improve the cell abundance estimation accuracy, classification results are corrected based on probabilities of interclass misclassification. The developed algorithms were tested using 158 phytoplankton images. There were 920 P. minimum cells in the images in total. The algorithms detected 98.1% of P. minimum cells present in the images and correctly classified 98.09% of all detected objects. The classification accuracy of detected P. minimum cells was equal to 98.9%, yielding a 97.0% overall recognition rate of P. minimum cells. The feature set used in this work has shown considerable tolerance to out-of-focus distortions. Tests of the system by phytoplankton experts in the cell abundance estimation task of P. minimum species have shown that its performance is comparable or even better than performance of phytoplankton experts exhibited in manual counting of artificial microparticles, similar to P. minimum cells. The automated system detected and correctly recognized 308 (91.1%) of 338 P. minimum cells found by experts in 65 phytoplankton images taken from new phytoplankton samples and erroneously assigned to the P. minimum class 3% of other objects. Note that, due to large variations of texture and size of P. minimum cells as well as background, the task performed by the system was more complex than that performed by the experts when counting artificial microparticles similar to P. minimum cells.

Fusion of voice signal information for detection of mild laryngeal pathology

Detection of mild laryngeal disorders using acoustic parameters of human voice is the main objective in this study. Observations of sustained phonation (audio recordings of vocalized /a/) are labeled by clinical diagnosis and rated by severity (from 0 to 3). Research is exclusively constrained to healthy (severity 0) and mildly pathological (severity 1) cases - two the most difficult classes to distinguish between. Comprehensive voice signal characterization and information fusion constitute the approach adopted here. Characterization is obtained through diverse feature set, containing 26 feature subsets of varying size, extracted from the voice signal. Usefulness of feature-level and decision-level fusion is explored using support vector machine (SVM) and random forest (RF) as basic classifiers. For both types of fusion we also investigate the influence of feature selection on model accuracy. To improve the decision-level fusion we introduce a simple unsupervised technique for ensemble design, which is based on partitioning the feature set by k-means clustering, where the parameter k controls the size and diversity of the prospective ensemble. All types of the fusion resulted in an evident improvement over the best individual feature subset. However, none of the types, including fusion setups comprising feature selection, proved to be significantly superior over the rest. The proposed ensemble design by feature set decomposition discernibly enhanced decision-level and significantly outperformed feature-level fusion. Ensemble of RF classifiers, induced from a cluster-based partitioning of the feature set, achieved equal error rate of 13.1+/-1.8% in the detection of mildly pathological larynx. This is a very encouraging result, considering that detection of mild laryngeal disorder is a more challenging task than a common discrimination between healthy and a wide spectrum of pathological cases.

Electromyographic Patterns during Golf Swing: Activation Sequence Profiling and Prediction of Shot Effectiveness

This study analyzes muscle activity, recorded in an eight-channel electromyographic (EMG) signal stream, during the golf swing using a 7-iron club and exploits information extracted from EMG dynamics to predict the success of the resulting shot. Muscles of the arm and shoulder on both the left and right sides, namely flexor carpi radialis, extensor digitorum communis, rhomboideus and trapezius, are considered for 15 golf players (∼5 shots each). The method using Gaussian filtering is outlined for EMG onset time estimation in each channel and activation sequence profiling. Shots of each player revealed a persistent pattern of muscle activation. Profiles were plotted and insights with respect to player effectiveness were provided. Inspection of EMG dynamics revealed a pair of highest peaks in each channel as the hallmark of golf swing, and a custom application of peak detection for automatic extraction of swing segment was introduced. Various EMG features, encompassing 22 feature sets, were constructed. Feature sets were used individually and also in decision-level fusion for the prediction of shot effectiveness. The prediction of the target attribute, such as club head speed or ball carry distance, was investigated using random forest as the learner in detection and regression tasks. Detection evaluates the personal effectiveness of a shot with respect to the player-specific average, whereas regression estimates the value of target attribute, using EMG features as predictors. Fusion after decision optimization provided the best results: the equal error rate in detection was 24.3% for the speed and 31.7% for the distance; the mean absolute percentage error in regression was 3.2% for the speed and 6.4% for the distance. Proposed EMG feature sets were found to be useful, especially when used in combination. Rankings of feature sets indicated statistics for muscle activity in both the left and right body sides, correlation-based analysis of EMG dynamics and features derived from the properties of two highest peaks as important predictors of personal shot effectiveness. Activation sequence profiles helped in analyzing muscle orchestration during golf shot, exposing a specific avalanche pattern, but data from more players are needed for stronger conclusions. Results demonstrate that information arising from an EMG signal stream is useful for predicting golf shot success, in terms of club head speed and ball carry distance, with acceptable accuracy. Surface EMG data, collected with a goal to automatically evaluate golf player’s performance, enables wearable computing in the field of ambient intelligence and has potential to enhance exercising of a long carry distance drive.

Exploring similarity-based classification of larynx disorders from human voice

In this paper identification of laryngeal disorders using cepstral parameters of human voice is researched. Mel-frequency cepstral coefficients (MFCCs), extracted from audio recordings of patients voice, are further approximated, using various strategies (sampling, averaging, and clustering by Gaussian mixture model). The effectiveness of similarity-based classification techniques in categorizing such pre-processed data into normal voice, nodular, and diffuse vocal fold lesion classes is explored and schemes to combine binary decisions of support vector machines (SVMs) are evaluated. Most practiced RBF kernel was compared to several constructed custom kernels: (i) a sequence kernel, defined over a pair of matrices, rather than over a pair of vectors and calculating the kernelized principal angle (KPA) between subspaces; (ii) a simple supervector kernel using only means of patients GMM; (iii) two distance kernels, specifically tailored to exploit covariance matrices of GMM and using the approximation of the Kullback-Leibler divergence from the Monte-Carlo sampling (KL-MCS), and the Kullback-Leibler divergence combined with the Earth movers distance (KL-EMD) as similarity metrics. The sequence kernel and the distance kernels both outperformed the popular RBF kernel, but the difference is statistically significant only in the distance kernels case. When tested on voice recordings, collected from 410 subjects (130 normal voice, 140 diffuse, and 140 nodular vocal fold lesions), the KL-MCS kernel, using GMM with full covariance matrices, and the KL-EMD kernel, using GMM with diagonal covariance matrices, provided the best overall performance. In most cases, SVM reached higher accuracy than least squares SVM, except for common binary classification using distance kernels. The results indicate that features, modeled with GMM, and kernel methods, exploiting this information, is an interesting fusion of generative (probabilistic) and discriminative (hyperplane) models for similarity-based classification.

Data dependent random forest applied to screening for laryngeal disorders through analysis of sustained phonation: acoustic versus contact microphone.

Comprehensive evaluation of results obtained using acoustic and contact microphones in screening for laryngeal disorders through analysis of sustained phonation is the main objective of this study. Aiming to obtain a versatile characterization of voice samples recorded using microphones of both types, 14 different sets of features are extracted and used to build an accurate classifier to distinguish between normal and pathological cases. We propose a new, data dependent random forests-based, way to combine information available from the different feature sets. An approach to exploring data and decisions made by a random forest is also presented. Experimental investigations using a mixed gender database of 273 subjects have shown that the perceptual linear predictive cepstral coefficients (PLPCC) was the best feature set for both microphones. However, the linear predictive coefficients (LPC) and linear predictive cosine transform coefficients (LPCTC) exhibited good performance in the acoustic microphone case only. Models designed using the acoustic microphone data significantly outperformed the ones built using data recorded by the contact microphone. The contact microphone did not bring any additional information useful for the classification. The proposed data dependent random forest significantly outperformed the traditional random forest.

Fusing voice and query data for non-invasive detection of laryngeal disorders

Voice and query data are explored for the task of laryngeal disorders detection.Decision-level fusion by complete-case analysis is compared to imputation strategies.Query data outperform voice, fusion after iterative model-based imputation - the best.Human readable rules were extracted from the query data using affinity analysis. Topic of this study is exploration and fusion of non-invasive measurements for an accurate detection of pathological larynx. Measurements for human subject encompass answers to items of a specific survey and information extracted by the openSMILE toolkit from several audio recordings of sustained phonation (vowel /a/). Clinical diagnosis, assigned by medical specialist, is a target attribute distinguishing subject as healthy or pathological. Random forest (RF) is used here as a base-learner and also as a meta-learner for decision-level fusion. 5 RF classifiers, built separately on 3 variants of audio recording data (raw and after two types of voice activity detection) and 2 variants of questionnaire (with 9 and 26 questions) data, are fused selectively by finding out the best combination of all possible. Before fusion, due to presence of missing values in query modalities, several imputation techniques were evaluated besides the complete-case analysis by listwise deletion. Out-of-bag equal error rate (EER) was found to be higher for audio data and lower for query, but each variant was outperformed by the decision-level fusion. Fusion after listwise deletion provided EER of 4.84%, meanwhile imputation was found to improve detection slightly and helped to achieve EER of 4.55%. Variable importance, as permutation-based mean decrease in RF accuracy, was reported for query and audio data. Finally, regarding the noteworthy performance of the query data, 22 association rules (11 healthy + 11 pathological) using 17 questions were obtained for comprehensible insights.

Automatic detection and morphological delineation of bacteriophages in electron microscopy images

Automatic detection, recognition and geometric characterization of bacteriophages in electron microscopy images was the main objective of this work. A novel technique, combining phase congruency-based image enhancement, Hough transform-, Radon transform- and open active contours with free boundary conditions-based object detection was developed to detect and recognize the bacteriophages associated with infection and lysis of cyanobacteria Aphanizomenon flos-aquae. A random forest classifier designed to recognize phage capsids provided higher than 99% accuracy, while measurable phage tails were detected and associated with a correct capsid with 81.35% accuracy. Automatically derived morphometric measurements of phage capsids and tails exhibited lower variability than the ones obtained manually. The technique allows performing precise and accurate quantitative (e.g. abundance estimation) and qualitative (e.g. diversity and capsid size) measurements for studying the interactions between host population and different phages that infect the same host.

Learning Accurate Active Contours

Focus of research in Active contour models (ACM) area is mainly on development of various energy functions based on physical intuition. In this work, instead of designing a new energy function, we generate a multitude of contour candidates using various values of ACM parameters, assess their quality, and select the most suitable one for an object at hand. A random forest is trained to make contour quality assessments. We demonstrate experimentally superiority of the developed technique over three known algorithms in the P. minimum cells detection task solved via segmentation of phytoplankton images.

Boosting performance of the edge-based active contour model applied to phytoplankton images

Automated contour detection for objects representing the Prorocentrum minimum (P. minimum) species in phytoplankton images is the core goal of this study. The species is known to cause harmful blooms in many estuarine and coastal environments. Active contour model (ACM)-based image segmentation is the approach adopted here as a potential solution. Currently, the main research in ACM area is highly focused on development of various energy functions having some physical intuition. This work, by contrast, advocates the idea of rich and diverse image preprocessing before segmentation. Advantage of the proposed preprocessing is demonstrated experimentally by comparing it to the six well known active contour techniques applied to the cell segmentation in microscopy imagery task.

Detecting Parkinson's disease from sustained phonation and speech signals

This study investigates signals from sustained phonation and text-dependent speech modalities for Parkinson’s disease screening. Phonation corresponds to the vowel /a/ voicing task and speech to the pronunciation of a short sentence in Lithuanian language. Signals were recorded through two channels simultaneously, namely, acoustic cardioid (AC) and smart phone (SP) microphones. Additional modalities were obtained by splitting speech recording into voiced and unvoiced parts. Information in each modality is summarized by 18 well-known audio feature sets. Random forest (RF) is used as a machine learning algorithm, both for individual feature sets and for decision-level fusion. Detection performance is measured by the out-of-bag equal error rate (EER) and the cost of log-likelihood-ratio. Essentia audio feature set was the best using the AC speech modality and YAAFE audio feature set was the best using the SP unvoiced modality, achieving EER of 20.30% and 25.57%, respectively. Fusion of all feature sets and modalities resulted in EER of 19.27% for the AC and 23.00% for the SP channel. Non-linear projection of a RF-based proximity matrix into the 2D space enriched medical decision support by visualization.